Date and Time
May 27, 2009
1:30 PM - 3:00 PM

Availability
Open to the public
No RSVP required

Speaker
Stefanos A. Zenios - Stanford University

The semi-parametric Cox proportional hazards model is commonly used to model the effect of various covariates on survival outcomes. However, its underlying proportional hazards assumption is not always appropriate. We introduce a nonparametric survival model for censored data, which constructs Kaplan-Meier survival curves based on the observed survival of the K-nearest neighbors obtained using Mahalanobis distance as the distance metric. K was selected to maximize the area under the Receiver-Operator Characteristic (ROC) curve calculated at a fixed time interval in a holdout dataset. Our method has the advantage of making no structural assumptions, and therefore can be applied to settings where the proportionality assumption is violated. We test the method in two settings. First, we apply it to a simulated dataset in which the proportional hazard assumption is clearly violated, and confirm that our method outperforms the standard Cox model. Second, we apply it to the setting of kidney transplantation to generate patient-specific survival distributions at the time of kidney transplantation or addition to the transplant waitlist. Separate transplant forecasts are generated for three different donor types (living, standard criteria, expanded criteria) and two endpoints (graft survival, patient survival). We test performance by comparing areas under the ROC curve against two alternative model formulations: A standard Cox model, and a hybrid K-nearest neighbors model that uses the difference in Cox score as the distance metric. Our method performs comparably to the other two methods in predicting post-transplant patient and graft survival, while outperforming the Cox model in waitlist survival prediction.